DARPA built an AI to fact-check enemy weapons claims

In late 2022, Chinese researchers claimed something extraordinary: Using today’s relatively basic quantum computers, they could hypothetically unlock encrypted information, the kind you could send from the messaging app Signal or spy satellites could beam to the ground. Coded, this information remains secure. But if quantum computers could crack this code, the secrets wouldn’t stay secret. Obviously, this is bad news for spicy Signal discussions, but it would be worse news for the defense community.

In fact, according to experts, this development would be disastrous, if it were real. But some of these same experts were skeptical. Was it hype? Fulminate? Regardless, this claim has raised concerns that China’s quantum and decryption capabilities at least could have overtaken the United States.

This type of fear does its own geopolitical dirty work, driving the United States to chase scientific shadows—a distraction the military can ill afford as it wages a war of choice with Iran.

On supporting science journalism

If you enjoy this article, please consider supporting our award-winning journalism by subscription. By purchasing a subscription, you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

Nobody wants that, especially not the nation’s premier military R&D organization, the Defense Advanced Research Projects Agency (DARPA). To assess the veracity of scientific developments, he launched SciFy, the scientific feasibility program. Currently in its first demonstrations, the program is building tools that can ingest a wild scientific claim and quickly dismiss it as BS — or flag it as a breakthrough that the Department of Defense should pay attention to.

DARPA’s job, says SciFy program manager Erica Briscoe, is to avoid technological surprises by ensuring that U.S. military technology always stays ahead of anything else out there. In fact, the agency was founded after the USSR surprise launched its first satellite, Sputnik, so that there would never be another Sputnik.

So in a way, SciFy is a sort of canonical program for DARPA. If its software tools work, they could assess whether rumors of a proverbial Sputnik are greatly exaggerated. “Feasibility goes beyond the commonly understood concepts of validation and replication,” Briscoe says, “and really gets into this speculative space that is a little bit of judgment and a little bit of art.”

SciFy’s tools can also allow defense and intelligence agencies to predict whether and by what steps another country will couldsay, develop a decryptor quantum computer in five years. “We might not think they have it, but we want to know how they might get it,” Briscoe says.

Fear of the adversary’s decryption capabilities does its own geopolitical dirty work.

AI-based analysis can also be prescriptive, revealing where to invest military R&D dollars – something DARPA is interested in. “If you’re an organization like that, you’re just trying to figure out, among all the range of crazy ideas that are out there in the world, what’s going to be valuable,” says Clayton Kerce of the Georgia Tech Research Institute, who is part of a SciFy team. In other words, beyond calling BS, SciFy tools could greenlight domestic projects.

Kerce’s team is working toward this goal by creating a toolset called Farscape. Right now, he says, you can input a scientific claim, and the system will spawn AI agents that will gather relevant information, reason about that body of knowledge, rank the importance of the evidence, and construct a BS or green light rating for the original claim. Farscape compares the results of these agents, summarizes them and gives an overall score.

The analysis involves, in part, Farscape’s attempts to reason using thought patterns similar to those used by humans, such as deduction and induction. Reasoning is one of Kerce’s interests as an applied mathematician, and it makes Farscape field-agnostic: it doesn’t care whether you ask about immortal batteries or impervious armor. As an internal test, the team asked him to determine when Chinese chipmakers would be as good as America’s Nvidia.

Computer scientist Frank Ferraro of the University of Maryland, Baltimore County, a member of another SciFy team, considers the system his group is working on one of his hobbies: woodworking. “We approached it that way, at least from my perspective,” he says, “in terms of balancing building your store, building the tools that you have, and then learning the best way to use them.”

This team has over 25 people who build their own tools. The overall system (the woodshop) can then decide which tools are most useful in responding to a specific complaint. A tool, such as a hammer, breaks down a claim into its verifiable elements. For example, what would happen if rival country A claimed to have made armor from a material that could repair itself? For this to be true, the material would have to be strong in equatorial jungles as well as polar tundra. But perhaps the scientific literature or a simulation of the material’s properties indicates that it would be a liquid on a tropical summer day. No liquid armor allowed; therefore, infeasible.

Currently, DARPA, in collaboration with the Johns Hopkins Applied Physics Laboratory, is evaluating the teams’ work, putting it through three “technical sprints.” In the first, teams had 48 hours for their AI to evaluate scientific claims. Their results – and the reasoning behind them – were compared to analyzes from human experts in the respective fields, a process that will be repeated in the final two sprints.

DARPA recently completed its materials science sprint and is in the process of evaluating the results of an AI sprint. After that will come quantum computing – all topics, Briscoe says, that the Department of Defense is concerned about.

The statements that the groups were asked to evaluate may seem a bit boring but have military relevance. For example, it reads “additives containing fluorine in a liquid electrolyte allow Li-ion batteries to operate at up to 10 V,” meaning that adding a little something extra to a battery could, for example, make drones lighter and able to fly further.

In the recently completed first sprint, all teams reached “moderate” agreement with human materials science experts, which was DARPA’s goal. Additionally, Kerce says, AI analysis sometimes causes experts to reconsider their views. “They changed their assessment in 19% of cases,” he recalls. (DARPA is putting together the official numbers for the first sprint), because the AI ​​was able to connect the dots between huge volumes of information that humans can’t hold in their heads simultaneously.

In the coming months, these teams will use their AI to tackle quantum computing claims, potentially similar to those of Chinese researchers — and, later, these same types of tools could tell the U.S. military how to develop its own quantum tricks without venturing into unproductive sci-fi land.